A collaborative multi-agent architecture with physics-consistency validation for robust microseismic signal classification

Reliable classification of microseismic (MS) records is essential for geohazard early warning in deep underground construction. Corrupted or missing local observations can distort shared multi-sensor representations, and waveform similarity alone is often insufficient to separate engineering disturbances from MS events. This study proposes Physics-Informed Multi-Agent Reinforcement Learning (PI-MARL), a sensor-bound framework for classifying MS events, blasting signals, and interference signals. Each sensor channel is processed by an independent agent comprising a lightweight multi-scale 1D-CNN and a dueling deep Q-network. Agents exchange latent representations through hierarchical spatio-temporal attention organized by front and back sensor groups. A separate physics-consistency module validates preliminary MS proposals using inter-group arrival-time and characteristic-band spectral conditions when both groups yield valid observations; in single-group operation, only the spectral condition is applied. Inter-group energy distribution serves solely as supplementary propagation evidence and does not by itself determine acceptance. Training combines the task reward with a fixed penalty for physics-violating proposals and attention-weighted credit assignment. Over ten runs on a held-out test subset from an active deep-tunnel project, PI-MARL attained a mean (± standard deviation) macro F1-score of 97.48 ± 0.29%, macro recall of 97.23 ± 0.38%, and false alarm rate (FAR) of 0.81 ± 0.28%. FAR increased to 10.0% in the lowest evaluated signal-to-noise ratio interval, [10, 15) dB. With four of the eight sensors unavailable, FAR was approximately 2.5% when failures were randomly distributed and both groups retained valid observations, and 6.8% under complete front-group failure, where the framework ran in single-group mode. Relative to the evaluated baselines, degradation was more gradual as noise and sensor loss increased, indicating that the coordinated architecture remains effective under the engineering conditions examined here.